Data Understanding

Before trying to build a classifier we first need to do some exploration on and understanding of the data. In the first place we need to decide which information we are going to use for the classifier. For example the genre for each artist probably will be similar and thus will not be useful data. Furthermore we have two possible sets of features we can use to train the classifier with:

• Track Features, these features are returned by the get_track_audio_features() method of the spotifyr package. This method is also used to get the track features in the get_artist_audio_features() and get_album_audio_features() methods. These features are values that say something about the song in a whole, thus we will get 1 feature value per song.
• Track Analyis, these features are obtained using the get_track_audio_analysis() method fromt the spotifyr package. The analysis features are quite a bit more extensive than the track features, thus will probably contain a lot more information about the song. However this means more data, which will take up more disk space, take longer to obtain from Spotify, make the classifiation training take more time and make the model quite a bit more complex (since we now need to add a time dimension to our model).

Because of the reasons described above I’m first going to focus on creating a model created with the track features. If I fail to create a good model with these features I’m going to take a look at the track analysis features.

The track features include many features including the following numeric features that may be useful:

• Danceability
• Energy
• Key
• Loudness
• Mode
• Speechiness
• Accousticness
• Instrumentalness
• Liveness
• Valence
• Tempo

This is a lot of data in which some features may be very similar for all songs. It is useless to include this data in the trainingsdata for the classifier since it wouldn’t provide good information to distinguish two songs from each other, let alone different artists.

To give a good insight in these features and get a quick overview of which of these may be show some clear differences between the artists, I have combined all songs from all artists into one dataset. I have plotted each of them here (I would have made them interactive however due to limitations of the plotly package they are just images. From these plots you can observe that the Mode feature is not useful. Some other features don’t seem to show any clear patterns on their own too. That is why I have plotted each feature relative to the other features. If one of these plots already show clusters we probably only need to use these two features to train a classifier with. However I have put these plots in a separate document because of the amount (121) and because no plot seems to show any clear clusters.

So although the violinplots did seem to show some nice patterns, these patterns seem to be very similar among the artists. This thus means we need to perform Principal Component Analysis. I will elaborate more on this in the Data Preparation section. If the PCA provides us with good clusters we know that we can quite easily build a classifier, however if the PCA doesn’t provide us with any noticable clusters it may be possible that the data still can be clustered, but in higher dimensions. This, however, is quite hard to visualize, thus then I will probably just feed the data to the classifier and hopefully it will be able to draw relations between the features and the artist.

Dimensionality Reduction

Before we can feed the data to the classifier we first need to prepare the data. One part of data preparation is data reduction. This means that we reduce the initial dataset to be only data we are going to use for the classifier. Since we are going to use the track features we can all discard all data other than these features. I combined this data only into a new data frame and saved that as my new corpus. Previously I mentioned removing the Mode feature from our data as well, however since we are now going to perform PCA I will keep this feature for now. Of course we still need to include the Artist in our reduced dataset since we need to use that data as classes for the model.

Principal Component Analysis

As mentioned in the Data Understanding we need to apply Principal Component Analysis on the data since the features on their own or relative to one other feature didn’t show any good clusters. Principal component analysis means that we reduce the data to a new dataset where each column is an information rich column that captures as much possible variation from the initial data. This data may be even better to use than the features on their own since the PCA data will be more dense in information, and will contain only the relevant parts of the features. Since the principal components will take the differences between values we need to make sure to scale the data first. This is because for example the tempo will otherwise be the leading variable in the principal components because due to its big values (relative to the other features) it also encapsulates the most variance. After the scaling I made a PCA of the data, I have plotted the first two against each other (see Figure 1).

As you can see, the data is now seperated into two columns over the first two principal components. Although there are two columns and the data seems to be seperated this way, each artist looks to have around the same amount of songs per column, thus not being very helpful.
It however may be the case that another separation can be found in other components, since we have 11 of them. To see wether this is the case, if have plotted the first 3 components with k-means clustering applied in the story Further analysis on the Principal components.

K-Means Clustering

We can use clustering for this and k-means clustering in particular. This clustering algorithm first assigns each data point to a random cluster. Then it iteratively calculates the centerpoint of each cluster and assigns each data point to the cluster with the nearest centerpoint, this is done using some distance formula that handles multidimensional data (like Euclidean distance, Manhattan distance). To compare the standard features with the principal components I have applied k-means clustering on both the features data and principal components data with 7 clusters (we have 7 artists).

If the data can be effectively clustered we would see 7 separate clusters, and in each cluster there will be only songs of one artist.
However, as you can see in Figure 2 and Figure 3 (can be found interactive in the next two stories) most data points are all in one big cluster and the outlines of each cluster all cross each other. Even the principal components, which spread the data over two columns on the first two components has songs in each column in the same cluster. To further illustrate this I show below that cluster 1 contains songs of each artist.

We do see that the standard track features have some outliers in general, while the principal components only have small outliers in each column. It may however be possible that there is more spread in higher dimensions of the principal components, which I explored in the Further analysis on the Principal Components section.

Principal Components

The Principal Components are very abstract, it is therefore very hard to explain what part of the data contributed to any good or bad results when using the principal components for, for example, classification.

Principal components are designed to represent as much variation in the data as possible thus it will not represent each feature anymore; each component will probably represent a comination of them, with the first components probably being a combination of features with high variation.
To give more insight in what kind of music is represented by each value of the first 3 principal components I have made a 3 dimensional plot in which I have plotted the songs colored by their artist on the first 3 components in the x, y and z axis respectively.
In this plot I have included the clusters as well, but it can be noticed straight away that these clusters are all over the place, however the clusters seem to be the most separated on the third principal component (z-axis).
To get a better view of the songs I therefore suggest you to disable the clusters by either clicking once on the cluster part of the legend or double clicking on the artist part of the legend.

After creating this plot I have used it to pick songs on the limits of each component to listen to and compare. Using this comparison I gained insight in what type of song is represented by each principal component.

Principal Component 1

• Low values: Better flowing music, easier to listen to.
• High values: Very spiky aggresive music.

The first principal component seems to characterize the flow of a song. Higher values seem to correspond with very ‘spiky’ songs; hard bass and no real flow.
I had expected that the danceability feature would be very low for these songs, however a few samples did not show that either high or low values of the first principal component corresponded with any high or low values of danceability.

Principal Component 2

• Low values: low Instrumentalness, high Speechiness
• High values: high Instrumentalness, low Speechiness

The second principal component was quite noticable straight away, songs with low values had lots of speech in them, songs with high values had no or barely noticable voice in them.
To confirm my observation I have picked 4 songs per column; 2 with low values for the second principal component and 2 songs with high values for the second principal component. Below I made two simple tables in which you can clearly see the correspondence between the value of the second principal component and the speechiness and instrumentalness of the song.

Principal Component 3

• Low values: Very melodic, lots of synth
• High values: Very heavy bass,

As for the third principal component, clear differences between songs were already harder to notice. One big difference could however be noticed; Songs with low values for the third principal component had way more synth and thus melodic feel to them than songs with higher values for the third principal component. Some songs with high values for the third component didn’t even have some kind of chorus (which contains in most hardstyle songs most of the synth of the song) and were just 2 to 3 minutes of bass kick with some distorted noise.

Subsetting the data for the classifier

Since we are going to train a classifier we also need to seperate the data into two subsets:

• A trainingset, containing about 80% of the data. This data will be used to train the classifier.
• A testset, containing the remaining 20% of the data. This data will be used to test the classifier.

However, since taking only 20% of the data as validation data, this means we get only 140 songs to validate the classifier. In the best scenario this will mean 20 songs per artist, however since we don’t have an equal amount of songs per artist we will most likely not get 20 songs per artist in the validation data.

K-Fold Cross Validation

For this reason I’m going to perform k-fold cross validation on the data with 5 folds. This means that I’m going to shuffle the data and then divide the data into 5 parts (thus each part is 20% of the data). After that I will repeatedly take one part as testset and the other parts as trainingset. This way all data will be once testdata and 4 times trainingsdata, resulting in a model that has seen more data and thus is less overfitted on that data, giving a better estimate of the real performance of the model.

To do this I first shuffle the indices of the data. Then I create a vector with equal amount of items per fold. Lastly I assign to each index a fold.

set.seed(123)   # Set a seed to make reproducible results
k <- 5          # Amount of folds
shuffled_indices <- sample(1:nrow(all_songs))                          # Shuffle the indices
fold_breaks <- cut(1:length(shuffled_indices), breaks=k, labels=FALSE)  # Split the data into k equal parts
fold_indices <- fold_breaks[shuffled_indices]                           # Assign a fold to each index

Preparing the Track Features Data

The first data I am going to use to train a classifier are the track features. What data is included in this can be found in The Data.

To make sure each feature has the same impact on the model each feature must be scaled to be within the same limits. I choose to scale the data between 0 and 1.

scale_01 <- function(x){(x-min(x))/(max(x)-min(x))}   # Define scale function
train_features <- apply(all_songs[,1:11], 2, scale_01)     # Scale track features data

Then I converted the artist names to a factor vector for the support vector machine and I created a one-hot-encoding matrix of the artist names for the Keras model. A one-hot-encoding matrix is just a row of zeros with one 1 at the position of the true class. Thus the one-hot-encoding vector for a song made by Noisecontrollers will contain only zeros and 1 one at the position of Noisecontrollers in the one-hot-encoding matrix like this:

[1] 0 0 0 0 1 0 0

Preparing the Principal Components Data

Preparing the Pincipal Components is very easy since one of the return values of the prcomp function returns a matrix with for each component for each song a value. We can then just use this matrix as input features for the Support Vector Machine and Keras Deep Learning Model.
Since the labels did not change (each song is still produced by its original artist) we can reuse this data; the one-hot-encoding matrix as target data for the Keras model and the factorised labels as target data for the SVM.
We also don’t need to scale the data since the data was already scaled before it was put into the principal component analysis, and the analysis itself scales the data too.

Preparing the _Delta Chroma, Delta Timbre and Delta Delta Timbre Data

train_timbre <- apply(all_songs_delta_timbre[,5:28], 2, scale_01)

In order to prepare the delta data I only needed to take the relevant data columns of the initial dataset with all audio analysis and rescale the data to be within the [0,1] range, since some deltas are negative.

Introduction

As mentioned before I’m going to train multiple models, trained on different subsets of the initial dataset:

• A model trained on all features mentioned in The Data (Data Understanding).
• A model trained on (a subset of) the Principal Components.
• A model trained on deltas of timbre and chroma and delta of timbre deltas.

Multiclass classification models can be made in different ways. Because I am very intrigued by deep learning, I had the intention from the start to make a deep learning model to use for classification. Deep learning comes in many packages but I already had heard of TensorFlow and especially its embedding in the Keras library. Keras originally is a deep learning library for the Python programming language, however I had seen a c# version as well. When doing Datacamp courses I noticed a course called Introduction to TensorFlow in R which included lessons for Keras in R. So I looked into the Keras package for R and found that it works the same as Keras for Python except for some methods having other names.

Multiclass classification models can, however, be made a lot easier than creating a whole Keras model, namely with Support Vector Machines. Especially since R provides very easy tools to create one.

That is why I decided to not only create a Keras model but also a Support Vector Machine as comparison model to see whether deep learning is better practice for this kind of classification, or if it is way easier (with maybe better results) to just create a support vector machine model.

The Build Model Function

Since the original track features data has 11 columns we need to have 11 inputs in our network, however since I have multiple data sets I want to use, I made the input size variable.
Then I created two hidden layers with 25 and 15 nodes respectively. Each node uses relu as its activation function since this is the most basic activation function.

The relu activation function outputs, as can be seen in the picture, zero for all negative input and is linear for all positive input. Relu is generally seen as the go to activation function for simple neural networks. I may try different activation functions is the future.

The output layer consists of 7 nodes since we have 7 classes.
I have used the softmax activation for this output layer since we want for each class a probability. The softmax activation function makes sure this is the case by converting the outputs to all sum to one.

build_model <- function(input_size) {
  model <- keras_model_sequential() %>%
    layer_dense(units=25, activation='relu', input_shape = input_size) %>% # Fully connected input layer with 25 nodes
    layer_dense(units=15, activation='relu') %>%  # Second fully connected hidden layer with 15 nodes
    layer_dense(units=7, activation='softmax')    # Output layer with 7 nodes (7 classes)
  
  model %>% compile(
    optimizer = 'rmsprop',
    loss = 'categorical_crossentropy',
    metrics = c('accuracy')
  )
}

Then I created a similar function to the run_svm function described earlier to run the model in a k-fold corss validation manner on the provided data.
I have actually created two functions for this since I am using Tensorboard to better inspect the performance of the deep learning models.
However Tensorboard is designed to visualize the results of the model in real time and is therefore hosted on the localhost and can thus sadly not be included in this dashboard.
Instead I have downloaded the graphs of the accuracy on the training and validation data and show these instead with 0.9 smoothing applied to show a better trend of the accuracy over the epochs.
The arguments of these functions are very much self-explanatory and I will thus not go further in on this.

run_keras <- function(k, x, y, ind, num_epochs, b_size, FUN = build_model){
  all_scores <- c()
  
  for (i in 1:k) {
    cat("Processing fold #", i, "\n")
    
    # Prepare the validation data: data from partition #i
    inds <- which(ind == i, arr.ind = TRUE) 
    
    # Get the training and validation data
    val_x <- x[inds,]
    val_y <- y[inds,]
    train_x <- x[-inds,]
    train_y <- y[-inds,]
    
    # Build the Keras model
    model <- FUN(ncol(train_x))
    model %>% fit(train_x, train_y,
                  epochs=num_epochs, batch_size=b_size, verbose=0)
    
    # Evaluate the model on the validation data
    results <- model %>% evaluate(val_x, val_y, verbose = 0)
    print(paste("Loss:", results$loss, "Accuracy:", results$accuracy))
    all_scores <- c(all_scores, results$accuracy)
  }
  
  print(mean(all_scores))
}

run_keras_tensorboard <- function(k, x, y, ind, num_epochs, b_size, tensorboard_path, FUN = build_model){
  accuracy_histories <- c()
  
  for (i in 1:k) {
    cat("Processing fold #", i, "\n")
    
    # Prepare the validation data: data from partition #i
    inds <- which(ind == i, arr.ind = TRUE) 
    
    # Get the training and validation data
    val_x <- x[inds,]
    val_y <- y[inds,]
    train_x <- x[-inds,]
    train_y <- y[-inds,]
    
    # Build the Keras model
    model <- FUN(ncol(train_x))
    
    # Train the model (in silent mode, verbose=0)
    history <- model %>% fit(train_x, train_y,
                  epochs = num_epochs, batch_size = b_size, verbose = 0,
                  callbacks = callback_tensorboard(paste(tensorboard_path, "/run_", i, sep="")),
                  validation_data = list(val_x, val_y))
    
    accuracy_histories <- rbind(accuracy_histories, history$metrics$val_accuracy)
  } 
  
  # Run Tensorboard and create average accuracy dataframe
  tensorboard(tensorboard_path)
  average_accuracy_history <- data.frame(
    epoch=seq(1:ncol(accuracy_histories)),
    validation_accuracy=apply(accuracy_histories, 2, mean)
  )
  return(average_accuracy_history)
}

Results of a SVM on the Track Features Data

To assess the performance of each model I have created a confusion matrix of the results. I have put the accuracy above each confusion matrix. Each row gives the predicted label while each column stands for the actual label. This way we can see which artists were correctly classified most times and if an artist is classified incorrectly which other artist was choosen instead. If a model has very high accuracy we would see only zeros with on diagonal line of numbers going from top left to bottom right.

If we look at the accuracies we see that each model roughly has the same accuracy, small differences between the accuracies are probably explained by each fold having a different distribution of songs per artist. The mean accuracy of all models is 36%, which is, considering the type of data, not that bad.

When looking at the confusion matrices we see that most of the inaccuracy of the models are explained by all of them classifying most songs as being from Noisecontrollers. Considering this artist having most songs in my corpus begs me to question whether this model actually ‘learned’ patterns in the data or just found out that guessing Noisecontrollers most times was just the right answer. Especially if we look deeper into the confusion matrices we see that the second most predicted label is Headhunterz who have, after Noisecontrollers, the most songs in my corpus.

Results of a Keras Model on the Track Features Data

Again I used K-fold cross validation to train 5 different models on 4 parts of the data as training data and the other part as validation data.
During the training I kept track of the accuracies and loss values of the model on the training and validation data on each epoch. Because I kept track of these values I used more epochs than were probably needed to see what the best epoch amount would be.

As we can see the accuracy of the models on the trainingdata reached a peak around 160 epoch with a mean accuracy around 43%.
When looking at the accuracy of the models on the validation data some kind of peak can be distinguished here as well, hopefully indicating that the peak in the accuracies on the trainingsdata did not mean that the models were overfitted in the trainingdata. We do see however that the accuracies of the models on the validation data don’t show much of a trend and vary much more between the models than the training accuracies do.
I also wanted to create a confusion matrix of the outputs of the deep learning models, however due to the classes be converted to a one-hot-encoding matrix As a final test I once again trained the models on the data with 160 epochs and got a final accuracy around 30%.

Considering the time it took to train these models on 160 epochs (around 2 minutes) and lower accuracies I conclude that using a deep learning model on the track features data is not feasible.
It may be argued that using a SVM isn’t feasible too since the accuracy is still very low and you can see that the SVM just predicts the most common labels, to which common label is most probable for a certain song.

Results of a SVM on the Principal Components Data

I again have applied the same technique as for the track features data, however this time I have used the principal components data as training data.
Sadly to no avail, the accuracy is about the same, and again we see that the most common artists are preferred way over the other artists.
If the Keras model again provides the same results, something is probably going wrong somewhere and thus needs to be fixed.

Results of a Keras model on the Principal Components Data

Here too I have applied the same steps I performed when applying deep learning on the track features data.
And again too we see about the same results, which I was afraid of; The model got quite good results on the training data, however the validation accuracy is very bad. Each fold produces very dissimilar results albeit in roughly the same (very low) accuracy range.

What can be noticed is that the accuracy of the training data does not fluctuate as much as it did in the track features models. What is causing this is unclear to me, since I have used the same architecture, activation functions and optimizer for the models.

What is clear however is that something needs to change, since the data and architectures used now are not sufficient for creating good performing models.

Some insights

When doing some additional testing I found that adding more data gained better results. E.g. when using only the delta timbre and chroma I already gained better results than just the track features. When combining the delta timbre and chroma with the track features accuracy increased even further.
Because I had also extracted the delta of the delta timbre I have added this data my final data set as well.

The final dataset of all combined features now contained 47 columns:

• All Track Features
• Chroma Deltas
• Timbre Deltas (c1 … c12)
• Delta Timbre Deltas (d1 … d12)

The name of each datacolumn of the final corpus can be seen below.

 [1] "danceability"     "energy"           "key"             
 [4] "loudness"         "mode"             "speechiness"     
 [7] "acousticness"     "instrumentalness" "liveness"        
[10] "valence"          "tempo"            "C"               
[13] "C#|Db"            "D"                "D#|Eb"           
[16] "E"                "F"                "F#|Gb"           
[19] "G"                "G#|Ab"            "A"               
[22] "A#|Bb"            "B"                "c01"             
[25] "c02"              "c03"              "c04"             
[28] "c05"              "c06"              "c07"             
[31] "c08"              "c09"              "c10"             
[34] "c11"              "c12"              "d01"             
[37] "d02"              "d03"              "d04"             
[40] "d05"              "d06"              "d07"             
[43] "d08"              "d09"              "d10"             
[46] "d11"              "d12"             

Changing the model architecture

Since the data now contains more columns, the build_model function as described earlier did not work as well anymore. Although the input size was a variable passed as argument, the architecture of the network was based on the original dataset size: 11 columns. Since the data now contains 47 columns, the first layers (with fewer nodes) could be bottlenecking the performance of the model.

Therefore I have slightly adjusted the original build_model function to create build_model2. As can be seen not only did the size of the hidden layers increase, their activation is also changed. I did this after some feedback that suggested that an activation function which often just outputs 0 might not yield optimal results. A few quick tests further confirmed this.
I also tried the adam optimizer, which further increased the accuracy, so I sticked with that too.

build_model2 <- function(input_size) {
  model <- keras_model_sequential() %>%
    layer_dense(units=50, activation='elu', input_shape = input_size) %>% # Fully connected input layer with 50 nodes
    layer_dense(units=50, activation='elu') %>%  # Second fully connected hidden layer with 50 nodes
    layer_dense(units=7, activation='softmax')   # Output layer with 7 nodes (7 classes)
  
  model %>% compile(
    optimizer = 'adam',
    loss = 'categorical_crossentropy',
    metrics = c('accuracy')
  )
}

Sampling the data

As mentioned before I noticed that previous models just guessed the most common label when in doubt. This pattern was found in both the SVM and Keras model on both the track features as well as the principal components data.
I therefore decided to give each artist equal amount of songs in the corpus.

The amount of songs for each artist and its contribution to the total corpus can be seen in the first pie chart. We indeed see that both Noisecontrollers and Headhunterz are overrepresented in the corpus.

The minimum of songs for one artist was 46 (for D-Block & S-te-Fan), however subsampling all other artists to have only 45 songs in the corpus did not yield me very good results.
Instead I have used the mean amount of songs of each artist in the corpus, which is around 100. I therefore have subsampled all artists which have more than 100 songs in the corpus and oversampled all artists which have less than 100 songs in the corpus.
As can be seen in the bottom pie chart, each artist now has an equal amount of songs in the corpus.

Since the amount of songs has now changed I had to re-assign each song to a fold, which I have done below.
The oversampled_all_songs dataset contains all features of the over- and subsampled songs, as described above.

set.seed(123)
k <- 5
oversampled_shuffled <- sample(1:nrow(oversampled_all_songs))                      # Shuffle the indices
oversampled_breaks <- cut(1:length(oversampled_shuffled), breaks=k, labels=FALSE)  # Split the data into k equal parts
oversampled_indices <- oversampled_breaks[oversampled_shuffled]                    # Assign a fold to each index

Finally Good Results

As can be seen from the accuracies per fold and the mean accuracy over all folds, I was finally able to create a models that performed reasonably accurate.
Especially the confusion matrices now follow the pattern you would like to see in a confusion matrix; A diagonal line of high numbers going from top left to bottom right.

When evaluating these confusion matrices we see that in each fold about 1 or 2 artists where quite hard to classify, despite each artist being represented quite even in each fold. Interestingly, each fold different artists where badly classified, with Showtek being the artist with which the SVM models had most often hard times.
Due to each fold having 2 artists that were quite hard to indentify the overall accuracy of each fold quickly decreases, explaining the mean accuracy of 63%.

I am still very much pleased by the amount of increase in accuracy, especially since this accuracy isn’t explained by probabilities anymore.

Even better results

With a mean validation accuracy steadily above 70% after 500 epochs, I think I can claim that I have created a deep learning model that was finally able to outperform the support vector machine model equivalent trained on the same data.

It, however, took quite some testing and adjustments of especially the batch size, to get these results.
The batch size determines how many samples are passed through the network before it is updated. Lower batch size values mean more updates, which often means that the accuracy of the network increases more per epoch. The downside is that each epoch takes more time -updating a neural network uses gradient descent which takes a lot of time, even with optimizers- and causes the accuracy of the network to spike more. Higher batch size values result in less updates per epoch and thus less time spend per epoch, but causing less increase in accuracy. You therefore need to perform more epochs. A batch size value of 15 turned out to be quite good for me. As can be seen in the graphs I performed 750 epochs to see a good trend in accuracy. Due to have a batch size value of 15 we can also see that the graph behind the smoothed graph is alot smoother than the accuracy graphs of previous Keras models.

What is interesting and both worrying is that the accuracy of the model on the training data after 600 epochs reaches 100%. This means that the model is overfitted on the training data (thus is very good in only classifying only that part of the data). Overfitting most often results in the validation accuracy dropping from some point and the validation loss increasing heavily.
This first characteristic does not seem to occur, since the accuracy of the models on the validation accuracy keeps increasing, even after 600 epochs. We do however see that the validation loss even surpasses the starting loss after around 450 epochs. So despite the validation accuracy increase even after 600 epochs I have decided to take 300 epochs as number of epochs for the final model.

Novel Data

As mentioned earlier, I want to test my models and techniques on novel data; songs that have been released by one of the artists of my corpus after I made my initial corpus. A few songs have been released already. I have put some of them below:

The tempograms

When looking at the tempograms of the songs we see clear similarities between the tempograms of both songs of D-Block & S-te-Fan and clear differences between the songs of D-Block & S-te-Fan and the song of Da Tweekaz.

When looking at the songs of D-Block & S-te-Fan we see that it has a few segments in which there is a very clear tempo, while at segments where the tempo is more unclear the pattern between both songs of D-Block & S-te-Fan are similar. Also both songs of D-Block & S-te-Fan have the same dominant tempo over the course of the track.

This in contrast to Knockout which, as discussed earlier, follows a repeating pattern with first a tempo around 210 BPM, which rises to 300 BPM at the choruses. When the tempo is less clear the pattern of possible tempos is also very different to the pattern found in the songs of D-Block & S-te-Fan.

Creating the Final Models

Before we can classify we first need to create the final models, since during the k-fold cross validation I created a new model and discarded that model each fold.

final_svm <- svm(x=oversampled_full_data, y=oversampled_factors)

final_keras_model <- build_model2(ncol(oversampled_full_data))
final_keras_model %>% 
  fit(oversampled_full_data, oversampled_targets,
      epochs=300, batch_size=15, verbose=0)

Preparing some novel songs

I have taken the 3 novel songs that have been in the spotlight throughout the Exploration of Novel Data section as novel songs for the final test.
Since I only took the relavant data of each novel song for each story in the Exploration of Novel Data section, I first need to create the features for the novel songs that can be interpreted by the models.
This means that I have added the audio analysis to the original track features and extracted the delta timbre and chroma together with the delta of the delta timbre.

Then I have normalised the data features to be in the [0,1] range. However when doing this I have used the maximum and minimum values of the sampled data set to normalise the novel songs with respect to the training data.

The Final Test

Finally we can test the novel songs on the final models.

The Support Vector Machine Results

As we can sadly see, the Support Vector Machine was not able to correctly classify the artist of the 3 novel songs I gave it. Both songs from D-Block & S-te-Fan were classified as being produced by Brennan Heart and Sub Zero Project respectively, while the one novel song from Da Tweekaz was also classified as being produced by Sub Zero Project.

It is interesting that as song by D-Block & S-te-Fan is classified as being from Brennan Heart since most songs by Brennan Heart contain a lot of singing, this is actually a bit less in most of the songs of D-Block & S-te-Fan, however the two novel songs did contain quite some singing and high tones as described in Exploration of Novel Data. The fact that the other song from D-Block & S-te-Fan was classified as being produced by Sub Zero Project is also quite interesting since both novel songs from D-Block & S-te-Fan are produced at the same time as when they had a collaboration with Sub Zero Project and produced a few songs together. The type of this songs is quite similar as the type of songs they produced together.

I do find Da Tweekaz and Sub Zero Project very dissimilar. Especially with this song from Da Tweekaz that is quite heavily in the more general dance genre, while Sub Zero Project does make more hardcore / heavy-bass / less dancy songs. I therefore fail to see any reasons why this songs is classified wrong.

The Keras Model Results

Since the Keras model outputs the activations of each output node, I thought it would be more interesting to display these instead of taking the argmax of these outputs.
As can be seen is that both songs from D-Block & S-te-Fan are classified as being from Sub Zero Project, again I think the same explanation can be applied as I gave above, and therefore I do not think it is that much of a misclassification. However we do see that the activation for the true label is quite low, especially for the second song of D-Block & S-te-Fan, which has only 1% chance of being from D-Block & S-te-Fan. I will explain this further in the Final Conclusion.

The fact that the song from Da Tweekaz is classified as being from Headhunterz is also not a very big misclassification, although both artists do make very unqiue songs. Also the activation for the true label again is very low: 2% chance that the song from Da Tweekaz is actually made by Da Tweekaz.